ntb_transport.c

   1/*
   2 * This file is provided under a dual BSD/GPLv2 license.  When using or
   3 *   redistributing this file, you may do so under either license.
   4 *
   5 *   GPL LICENSE SUMMARY
   6 *
   7 *   Copyright(c) 2012 Intel Corporation. All rights reserved.
   8 *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
   9 *
  10 *   This program is free software; you can redistribute it and/or modify
  11 *   it under the terms of version 2 of the GNU General Public License as
  12 *   published by the Free Software Foundation.
  13 *
  14 *   BSD LICENSE
  15 *
  16 *   Copyright(c) 2012 Intel Corporation. All rights reserved.
  17 *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
  18 *
  19 *   Redistribution and use in source and binary forms, with or without
  20 *   modification, are permitted provided that the following conditions
  21 *   are met:
  22 *
  23 *     * Redistributions of source code must retain the above copyright
  24 *       notice, this list of conditions and the following disclaimer.
  25 *     * Redistributions in binary form must reproduce the above copy
  26 *       notice, this list of conditions and the following disclaimer in
  27 *       the documentation and/or other materials provided with the
  28 *       distribution.
  29 *     * Neither the name of Intel Corporation nor the names of its
  30 *       contributors may be used to endorse or promote products derived
  31 *       from this software without specific prior written permission.
  32 *
  33 *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  34 *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  35 *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  36 *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  37 *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  38 *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  39 *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  40 *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  41 *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  42 *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  43 *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  44 *
  45 * PCIe NTB Transport Linux driver
  46 *
  47 * Contact Information:
  48 * Jon Mason <jon.mason@intel.com>
  49 */
  50#include <linux/debugfs.h>
  51#include <linux/delay.h>
  52#include <linux/dmaengine.h>
  53#include <linux/dma-mapping.h>
  54#include <linux/errno.h>
  55#include <linux/export.h>
  56#include <linux/interrupt.h>
  57#include <linux/module.h>
  58#include <linux/pci.h>
  59#include <linux/slab.h>
  60#include <linux/types.h>
  61#include <linux/uaccess.h>
  62#include <linux/mutex.h>
  63#include "linux/ntb.h"
  64#include "linux/ntb_transport.h"
  65
  66#define NTB_TRANSPORT_VERSION	4
  67#define NTB_TRANSPORT_VER	"4"
  68#define NTB_TRANSPORT_NAME	"ntb_transport"
  69#define NTB_TRANSPORT_DESC	"Software Queue-Pair Transport over NTB"
  70#define NTB_TRANSPORT_MIN_SPADS (MW0_SZ_HIGH + 2)
  71
  72MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
  73MODULE_VERSION(NTB_TRANSPORT_VER);
  74MODULE_LICENSE("Dual BSD/GPL");
  75MODULE_AUTHOR("Intel Corporation");
  76
  77static unsigned long max_mw_size;
  78module_param(max_mw_size, ulong, 0644);
  79MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
  80
  81static unsigned int transport_mtu = 0x10000;
  82module_param(transport_mtu, uint, 0644);
  83MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  84
  85static unsigned char max_num_clients;
  86module_param(max_num_clients, byte, 0644);
  87MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  88
  89static unsigned int copy_bytes = 1024;
  90module_param(copy_bytes, uint, 0644);
  91MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  92
  93static bool use_dma;
  94module_param(use_dma, bool, 0644);
  95MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
  96
  97static bool use_msi;
  98#ifdef CONFIG_NTB_MSI
  99module_param(use_msi, bool, 0644);
 100MODULE_PARM_DESC(use_msi, "Use MSI interrupts instead of doorbells");
 101#endif
 102
 103static struct dentry *nt_debugfs_dir;
 104
 105/* Only two-ports NTB devices are supported */
 106#define PIDX		NTB_DEF_PEER_IDX
 107
 108struct ntb_queue_entry {
 109	/* ntb_queue list reference */
 110	struct list_head entry;
 111	/* pointers to data to be transferred */
 112	void *cb_data;
 113	void *buf;
 114	unsigned int len;
 115	unsigned int flags;
 116	int retries;
 117	int errors;
 118	unsigned int tx_index;
 119	unsigned int rx_index;
 120
 121	struct ntb_transport_qp *qp;
 122	union {
 123		struct ntb_payload_header __iomem *tx_hdr;
 124		struct ntb_payload_header *rx_hdr;
 125	};
 126};
 127
 128struct ntb_rx_info {
 129	unsigned int entry;
 130};
 131
 132struct ntb_transport_qp {
 133	struct ntb_transport_ctx *transport;
 134	struct ntb_dev *ndev;
 135	void *cb_data;
 136	struct dma_chan *tx_dma_chan;
 137	struct dma_chan *rx_dma_chan;
 138
 139	bool client_ready;
 140	bool link_is_up;
 141	bool active;
 142
 143	u8 qp_num;	/* Only 64 QP's are allowed.  0-63 */
 144	u64 qp_bit;
 145
 146	struct ntb_rx_info __iomem *rx_info;
 147	struct ntb_rx_info *remote_rx_info;
 148
 149	void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 150			   void *data, int len);
 151	struct list_head tx_free_q;
 152	spinlock_t ntb_tx_free_q_lock;
 153	void __iomem *tx_mw;
 154	phys_addr_t tx_mw_phys;
 155	size_t tx_mw_size;
 156	dma_addr_t tx_mw_dma_addr;
 157	unsigned int tx_index;
 158	unsigned int tx_max_entry;
 159	unsigned int tx_max_frame;
 160
 161	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 162			   void *data, int len);
 163	struct list_head rx_post_q;
 164	struct list_head rx_pend_q;
 165	struct list_head rx_free_q;
 166	/* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
 167	spinlock_t ntb_rx_q_lock;
 168	void *rx_buff;
 169	unsigned int rx_index;
 170	unsigned int rx_max_entry;
 171	unsigned int rx_max_frame;
 172	unsigned int rx_alloc_entry;
 173	dma_cookie_t last_cookie;
 174	struct tasklet_struct rxc_db_work;
 175
 176	void (*event_handler)(void *data, int status);
 177	struct delayed_work link_work;
 178	struct work_struct link_cleanup;
 179
 180	struct dentry *debugfs_dir;
 181	struct dentry *debugfs_stats;
 182
 183	/* Stats */
 184	u64 rx_bytes;
 185	u64 rx_pkts;
 186	u64 rx_ring_empty;
 187	u64 rx_err_no_buf;
 188	u64 rx_err_oflow;
 189	u64 rx_err_ver;
 190	u64 rx_memcpy;
 191	u64 rx_async;
 192	u64 tx_bytes;
 193	u64 tx_pkts;
 194	u64 tx_ring_full;
 195	u64 tx_err_no_buf;
 196	u64 tx_memcpy;
 197	u64 tx_async;
 198
 199	bool use_msi;
 200	int msi_irq;
 201	struct ntb_msi_desc msi_desc;
 202	struct ntb_msi_desc peer_msi_desc;
 203};
 204
 205struct ntb_transport_mw {
 206	phys_addr_t phys_addr;
 207	resource_size_t phys_size;
 208	void __iomem *vbase;
 209	size_t xlat_size;
 210	size_t buff_size;
 211	size_t alloc_size;
 212	void *alloc_addr;
 213	void *virt_addr;
 214	dma_addr_t dma_addr;
 215};
 216
 217struct ntb_transport_client_dev {
 218	struct list_head entry;
 219	struct ntb_transport_ctx *nt;
 220	struct device dev;
 221};
 222
 223struct ntb_transport_ctx {
 224	struct list_head entry;
 225	struct list_head client_devs;
 226
 227	struct ntb_dev *ndev;
 228
 229	struct ntb_transport_mw *mw_vec;
 230	struct ntb_transport_qp *qp_vec;
 231	unsigned int mw_count;
 232	unsigned int qp_count;
 233	u64 qp_bitmap;
 234	u64 qp_bitmap_free;
 235
 236	bool use_msi;
 237	unsigned int msi_spad_offset;
 238	u64 msi_db_mask;
 239
 240	bool link_is_up;
 241	struct delayed_work link_work;
 242	struct work_struct link_cleanup;
 243
 244	struct dentry *debugfs_node_dir;
 245
 246	/* Make sure workq of link event be executed serially */
 247	struct mutex link_event_lock;
 248};
 249
 250enum {
 251	DESC_DONE_FLAG = BIT(0),
 252	LINK_DOWN_FLAG = BIT(1),
 253};
 254
 255struct ntb_payload_header {
 256	unsigned int ver;
 257	unsigned int len;
 258	unsigned int flags;
 259};
 260
 261enum {
 262	VERSION = 0,
 263	QP_LINKS,
 264	NUM_QPS,
 265	NUM_MWS,
 266	MW0_SZ_HIGH,
 267	MW0_SZ_LOW,
 268};
 269
 270#define dev_client_dev(__dev) \
 271	container_of((__dev), struct ntb_transport_client_dev, dev)
 272
 273#define drv_client(__drv) \
 274	container_of((__drv), struct ntb_transport_client, driver)
 275
 276#define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
 277#define NTB_QP_DEF_NUM_ENTRIES	100
 278#define NTB_LINK_DOWN_TIMEOUT	10
 279
 280static void ntb_transport_rxc_db(unsigned long data);
 281static const struct ntb_ctx_ops ntb_transport_ops;
 282static struct ntb_client ntb_transport_client;
 283static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 284			       struct ntb_queue_entry *entry);
 285static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
 286static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
 287static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
 288
 289
 290static int ntb_transport_bus_match(struct device *dev,
 291				   const struct device_driver *drv)
 292{
 293	return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 294}
 295
 296static int ntb_transport_bus_probe(struct device *dev)
 297{
 298	const struct ntb_transport_client *client;
 299	int rc;
 300
 301	get_device(dev);
 302
 303	client = drv_client(dev->driver);
 304	rc = client->probe(dev);
 305	if (rc)
 306		put_device(dev);
 307
 308	return rc;
 309}
 310
 311static void ntb_transport_bus_remove(struct device *dev)
 312{
 313	const struct ntb_transport_client *client;
 314
 315	client = drv_client(dev->driver);
 316	client->remove(dev);
 317
 318	put_device(dev);
 319}
 320
 321static const struct bus_type ntb_transport_bus = {
 322	.name = "ntb_transport",
 323	.match = ntb_transport_bus_match,
 324	.probe = ntb_transport_bus_probe,
 325	.remove = ntb_transport_bus_remove,
 326};
 327
 328static LIST_HEAD(ntb_transport_list);
 329
 330static int ntb_bus_init(struct ntb_transport_ctx *nt)
 331{
 332	list_add_tail(&nt->entry, &ntb_transport_list);
 333	return 0;
 334}
 335
 336static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 337{
 338	struct ntb_transport_client_dev *client_dev, *cd;
 339
 340	list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 341		dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 342			dev_name(&client_dev->dev));
 343		list_del(&client_dev->entry);
 344		device_unregister(&client_dev->dev);
 345	}
 346
 347	list_del(&nt->entry);
 348}
 349
 350static void ntb_transport_client_release(struct device *dev)
 351{
 352	struct ntb_transport_client_dev *client_dev;
 353
 354	client_dev = dev_client_dev(dev);
 355	kfree(client_dev);
 356}
 357
 358/**
 359 * ntb_transport_unregister_client_dev - Unregister NTB client device
 360 * @device_name: Name of NTB client device
 361 *
 362 * Unregister an NTB client device with the NTB transport layer
 363 */
 364void ntb_transport_unregister_client_dev(char *device_name)
 365{
 366	struct ntb_transport_client_dev *client, *cd;
 367	struct ntb_transport_ctx *nt;
 368
 369	list_for_each_entry(nt, &ntb_transport_list, entry)
 370		list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 371			if (!strncmp(dev_name(&client->dev), device_name,
 372				     strlen(device_name))) {
 373				list_del(&client->entry);
 374				device_unregister(&client->dev);
 375			}
 376}
 377EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 378
 379/**
 380 * ntb_transport_register_client_dev - Register NTB client device
 381 * @device_name: Name of NTB client device
 382 *
 383 * Register an NTB client device with the NTB transport layer
 384 *
 385 * Returns: %0 on success or -errno code on error
 386 */
 387int ntb_transport_register_client_dev(char *device_name)
 388{
 389	struct ntb_transport_client_dev *client_dev;
 390	struct ntb_transport_ctx *nt;
 391	int node;
 392	int rc, i = 0;
 393
 394	if (list_empty(&ntb_transport_list))
 395		return -ENODEV;
 396
 397	list_for_each_entry(nt, &ntb_transport_list, entry) {
 398		struct device *dev;
 399
 400		node = dev_to_node(&nt->ndev->dev);
 401
 402		client_dev = kzalloc_node(sizeof(*client_dev),
 403					  GFP_KERNEL, node);
 404		if (!client_dev) {
 405			rc = -ENOMEM;
 406			goto err;
 407		}
 408
 409		dev = &client_dev->dev;
 410
 411		/* setup and register client devices */
 412		dev_set_name(dev, "%s%d", device_name, i);
 413		dev->bus = &ntb_transport_bus;
 414		dev->release = ntb_transport_client_release;
 415		dev->parent = &nt->ndev->dev;
 416
 417		rc = device_register(dev);
 418		if (rc) {
 419			put_device(dev);
 420			goto err;
 421		}
 422
 423		list_add_tail(&client_dev->entry, &nt->client_devs);
 424		i++;
 425	}
 426
 427	return 0;
 428
 429err:
 430	ntb_transport_unregister_client_dev(device_name);
 431
 432	return rc;
 433}
 434EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 435
 436/**
 437 * ntb_transport_register_client - Register NTB client driver
 438 * @drv: NTB client driver to be registered
 439 *
 440 * Register an NTB client driver with the NTB transport layer
 441 *
 442 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 443 */
 444int ntb_transport_register_client(struct ntb_transport_client *drv)
 445{
 446	drv->driver.bus = &ntb_transport_bus;
 447
 448	if (list_empty(&ntb_transport_list))
 449		return -ENODEV;
 450
 451	return driver_register(&drv->driver);
 452}
 453EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 454
 455/**
 456 * ntb_transport_unregister_client - Unregister NTB client driver
 457 * @drv: NTB client driver to be unregistered
 458 *
 459 * Unregister an NTB client driver with the NTB transport layer
 460 *
 461 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 462 */
 463void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 464{
 465	driver_unregister(&drv->driver);
 466}
 467EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 468
 469static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 470			    loff_t *offp)
 471{
 472	struct ntb_transport_qp *qp;
 473	char *buf;
 474	ssize_t ret, out_offset, out_count;
 475
 476	qp = filp->private_data;
 477
 478	if (!qp || !qp->link_is_up)
 479		return 0;
 480
 481	out_count = 1000;
 482
 483	buf = kmalloc(out_count, GFP_KERNEL);
 484	if (!buf)
 485		return -ENOMEM;
 486
 487	out_offset = 0;
 488	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 489			       "\nNTB QP stats:\n\n");
 490	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 491			       "rx_bytes - \t%llu\n", qp->rx_bytes);
 492	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 493			       "rx_pkts - \t%llu\n", qp->rx_pkts);
 494	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 495			       "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 496	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 497			       "rx_async - \t%llu\n", qp->rx_async);
 498	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 499			       "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 500	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 501			       "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 502	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 503			       "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 504	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 505			       "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 506	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 507			       "rx_buff - \t0x%p\n", qp->rx_buff);
 508	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 509			       "rx_index - \t%u\n", qp->rx_index);
 510	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 511			       "rx_max_entry - \t%u\n", qp->rx_max_entry);
 512	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 513			       "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
 514
 515	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 516			       "tx_bytes - \t%llu\n", qp->tx_bytes);
 517	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 518			       "tx_pkts - \t%llu\n", qp->tx_pkts);
 519	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 520			       "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 521	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 522			       "tx_async - \t%llu\n", qp->tx_async);
 523	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 524			       "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 525	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 526			       "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 527	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 528			       "tx_mw - \t0x%p\n", qp->tx_mw);
 529	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 530			       "tx_index (H) - \t%u\n", qp->tx_index);
 531	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 532			       "RRI (T) - \t%u\n",
 533			       qp->remote_rx_info->entry);
 534	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 535			       "tx_max_entry - \t%u\n", qp->tx_max_entry);
 536	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 537			       "free tx - \t%u\n",
 538			       ntb_transport_tx_free_entry(qp));
 539
 540	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 541			       "\n");
 542	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 543			       "Using TX DMA - \t%s\n",
 544			       qp->tx_dma_chan ? "Yes" : "No");
 545	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 546			       "Using RX DMA - \t%s\n",
 547			       qp->rx_dma_chan ? "Yes" : "No");
 548	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 549			       "QP Link - \t%s\n",
 550			       qp->link_is_up ? "Up" : "Down");
 551	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 552			       "\n");
 553
 554	if (out_offset > out_count)
 555		out_offset = out_count;
 556
 557	ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 558	kfree(buf);
 559	return ret;
 560}
 561
 562static const struct file_operations ntb_qp_debugfs_stats = {
 563	.owner = THIS_MODULE,
 564	.open = simple_open,
 565	.read = debugfs_read,
 566};
 567
 568static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 569			 struct list_head *list)
 570{
 571	unsigned long flags;
 572
 573	spin_lock_irqsave(lock, flags);
 574	list_add_tail(entry, list);
 575	spin_unlock_irqrestore(lock, flags);
 576}
 577
 578static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 579					   struct list_head *list)
 580{
 581	struct ntb_queue_entry *entry;
 582	unsigned long flags;
 583
 584	spin_lock_irqsave(lock, flags);
 585	if (list_empty(list)) {
 586		entry = NULL;
 587		goto out;
 588	}
 589	entry = list_first_entry(list, struct ntb_queue_entry, entry);
 590	list_del(&entry->entry);
 591
 592out:
 593	spin_unlock_irqrestore(lock, flags);
 594
 595	return entry;
 596}
 597
 598static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
 599					   struct list_head *list,
 600					   struct list_head *to_list)
 601{
 602	struct ntb_queue_entry *entry;
 603	unsigned long flags;
 604
 605	spin_lock_irqsave(lock, flags);
 606
 607	if (list_empty(list)) {
 608		entry = NULL;
 609	} else {
 610		entry = list_first_entry(list, struct ntb_queue_entry, entry);
 611		list_move_tail(&entry->entry, to_list);
 612	}
 613
 614	spin_unlock_irqrestore(lock, flags);
 615
 616	return entry;
 617}
 618
 619static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 620				     unsigned int qp_num)
 621{
 622	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 623	struct ntb_transport_mw *mw;
 624	struct ntb_dev *ndev = nt->ndev;
 625	struct ntb_queue_entry *entry;
 626	unsigned int rx_size, num_qps_mw;
 627	unsigned int mw_num, mw_count, qp_count;
 628	unsigned int i;
 629	int node;
 630
 631	mw_count = nt->mw_count;
 632	qp_count = nt->qp_count;
 633
 634	mw_num = QP_TO_MW(nt, qp_num);
 635	mw = &nt->mw_vec[mw_num];
 636
 637	if (!mw->virt_addr)
 638		return -ENOMEM;
 639
 640	if (mw_num < qp_count % mw_count)
 641		num_qps_mw = qp_count / mw_count + 1;
 642	else
 643		num_qps_mw = qp_count / mw_count;
 644
 645	rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 646	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
 647	rx_size -= sizeof(struct ntb_rx_info);
 648
 649	qp->remote_rx_info = qp->rx_buff + rx_size;
 650
 651	/* Due to housekeeping, there must be atleast 2 buffs */
 652	qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 653	qp->rx_max_entry = rx_size / qp->rx_max_frame;
 654	qp->rx_index = 0;
 655
 656	/*
 657	 * Checking to see if we have more entries than the default.
 658	 * We should add additional entries if that is the case so we
 659	 * can be in sync with the transport frames.
 660	 */
 661	node = dev_to_node(&ndev->dev);
 662	for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
 663		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
 664		if (!entry)
 665			return -ENOMEM;
 666
 667		entry->qp = qp;
 668		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
 669			     &qp->rx_free_q);
 670		qp->rx_alloc_entry++;
 671	}
 672
 673	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 674
 675	/* setup the hdr offsets with 0's */
 676	for (i = 0; i < qp->rx_max_entry; i++) {
 677		void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 678				sizeof(struct ntb_payload_header));
 679		memset(offset, 0, sizeof(struct ntb_payload_header));
 680	}
 681
 682	qp->rx_pkts = 0;
 683	qp->tx_pkts = 0;
 684	qp->tx_index = 0;
 685
 686	return 0;
 687}
 688
 689static irqreturn_t ntb_transport_isr(int irq, void *dev)
 690{
 691	struct ntb_transport_qp *qp = dev;
 692
 693	tasklet_schedule(&qp->rxc_db_work);
 694
 695	return IRQ_HANDLED;
 696}
 697
 698static void ntb_transport_setup_qp_peer_msi(struct ntb_transport_ctx *nt,
 699					    unsigned int qp_num)
 700{
 701	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 702	int spad = qp_num * 2 + nt->msi_spad_offset;
 703
 704	if (!nt->use_msi)
 705		return;
 706
 707	if (spad >= ntb_spad_count(nt->ndev))
 708		return;
 709
 710	qp->peer_msi_desc.addr_offset =
 711		ntb_peer_spad_read(qp->ndev, PIDX, spad);
 712	qp->peer_msi_desc.data =
 713		ntb_peer_spad_read(qp->ndev, PIDX, spad + 1);
 714
 715	dev_dbg(&qp->ndev->pdev->dev, "QP%d Peer MSI addr=%x data=%x\n",
 716		qp_num, qp->peer_msi_desc.addr_offset, qp->peer_msi_desc.data);
 717
 718	if (qp->peer_msi_desc.addr_offset) {
 719		qp->use_msi = true;
 720		dev_info(&qp->ndev->pdev->dev,
 721			 "Using MSI interrupts for QP%d\n", qp_num);
 722	}
 723}
 724
 725static void ntb_transport_setup_qp_msi(struct ntb_transport_ctx *nt,
 726				       unsigned int qp_num)
 727{
 728	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 729	int spad = qp_num * 2 + nt->msi_spad_offset;
 730	int rc;
 731
 732	if (!nt->use_msi)
 733		return;
 734
 735	if (spad >= ntb_spad_count(nt->ndev)) {
 736		dev_warn_once(&qp->ndev->pdev->dev,
 737			      "Not enough SPADS to use MSI interrupts\n");
 738		return;
 739	}
 740
 741	ntb_spad_write(qp->ndev, spad, 0);
 742	ntb_spad_write(qp->ndev, spad + 1, 0);
 743
 744	if (!qp->msi_irq) {
 745		qp->msi_irq = ntbm_msi_request_irq(qp->ndev, ntb_transport_isr,
 746						   KBUILD_MODNAME, qp,
 747						   &qp->msi_desc);
 748		if (qp->msi_irq < 0) {
 749			dev_warn(&qp->ndev->pdev->dev,
 750				 "Unable to allocate MSI interrupt for qp%d\n",
 751				 qp_num);
 752			return;
 753		}
 754	}
 755
 756	rc = ntb_spad_write(qp->ndev, spad, qp->msi_desc.addr_offset);
 757	if (rc)
 758		goto err_free_interrupt;
 759
 760	rc = ntb_spad_write(qp->ndev, spad + 1, qp->msi_desc.data);
 761	if (rc)
 762		goto err_free_interrupt;
 763
 764	dev_dbg(&qp->ndev->pdev->dev, "QP%d MSI %d addr=%x data=%x\n",
 765		qp_num, qp->msi_irq, qp->msi_desc.addr_offset,
 766		qp->msi_desc.data);
 767
 768	return;
 769
 770err_free_interrupt:
 771	devm_free_irq(&nt->ndev->dev, qp->msi_irq, qp);
 772}
 773
 774static void ntb_transport_msi_peer_desc_changed(struct ntb_transport_ctx *nt)
 775{
 776	int i;
 777
 778	dev_dbg(&nt->ndev->pdev->dev, "Peer MSI descriptors changed");
 779
 780	for (i = 0; i < nt->qp_count; i++)
 781		ntb_transport_setup_qp_peer_msi(nt, i);
 782}
 783
 784static void ntb_transport_msi_desc_changed(void *data)
 785{
 786	struct ntb_transport_ctx *nt = data;
 787	int i;
 788
 789	dev_dbg(&nt->ndev->pdev->dev, "MSI descriptors changed");
 790
 791	for (i = 0; i < nt->qp_count; i++)
 792		ntb_transport_setup_qp_msi(nt, i);
 793
 794	ntb_peer_db_set(nt->ndev, nt->msi_db_mask);
 795}
 796
 797static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 798{
 799	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 800	struct pci_dev *pdev = nt->ndev->pdev;
 801
 802	if (!mw->virt_addr)
 803		return;
 804
 805	ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
 806	dma_free_coherent(&pdev->dev, mw->alloc_size,
 807			  mw->alloc_addr, mw->dma_addr);
 808	mw->xlat_size = 0;
 809	mw->buff_size = 0;
 810	mw->alloc_size = 0;
 811	mw->alloc_addr = NULL;
 812	mw->virt_addr = NULL;
 813}
 814
 815static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
 816			       struct device *ntb_dev, size_t align)
 817{
 818	dma_addr_t dma_addr;
 819	void *alloc_addr, *virt_addr;
 820	int rc;
 821
 822	/*
 823	 * The buffer here is allocated against the NTB device. The reason to
 824	 * use dma_alloc_*() call is to allocate a large IOVA contiguous buffer
 825	 * backing the NTB BAR for the remote host to write to. During receive
 826	 * processing, the data is being copied out of the receive buffer to
 827	 * the kernel skbuff. When a DMA device is being used, dma_map_page()
 828	 * is called on the kvaddr of the receive buffer (from dma_alloc_*())
 829	 * and remapped against the DMA device. It appears to be a double
 830	 * DMA mapping of buffers, but first is mapped to the NTB device and
 831	 * second is to the DMA device. DMA_ATTR_FORCE_CONTIGUOUS is necessary
 832	 * in order for the later dma_map_page() to not fail.
 833	 */
 834	alloc_addr = dma_alloc_attrs(ntb_dev, mw->alloc_size,
 835				     &dma_addr, GFP_KERNEL,
 836				     DMA_ATTR_FORCE_CONTIGUOUS);
 837	if (!alloc_addr) {
 838		dev_err(ntb_dev, "Unable to alloc MW buff of size %zu\n",
 839			mw->alloc_size);
 840		return -ENOMEM;
 841	}
 842	virt_addr = alloc_addr;
 843
 844	/*
 845	 * we must ensure that the memory address allocated is BAR size
 846	 * aligned in order for the XLAT register to take the value. This
 847	 * is a requirement of the hardware. It is recommended to setup CMA
 848	 * for BAR sizes equal or greater than 4MB.
 849	 */
 850	if (!IS_ALIGNED(dma_addr, align)) {
 851		if (mw->alloc_size > mw->buff_size) {
 852			virt_addr = PTR_ALIGN(alloc_addr, align);
 853			dma_addr = ALIGN(dma_addr, align);
 854		} else {
 855			rc = -ENOMEM;
 856			goto err;
 857		}
 858	}
 859
 860	mw->alloc_addr = alloc_addr;
 861	mw->virt_addr = virt_addr;
 862	mw->dma_addr = dma_addr;
 863
 864	return 0;
 865
 866err:
 867	dma_free_coherent(ntb_dev, mw->alloc_size, alloc_addr, dma_addr);
 868
 869	return rc;
 870}
 871
 872static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 873		      resource_size_t size)
 874{
 875	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 876	struct pci_dev *pdev = nt->ndev->pdev;
 877	size_t xlat_size, buff_size;
 878	resource_size_t xlat_align;
 879	resource_size_t xlat_align_size;
 880	int rc;
 881
 882	if (!size)
 883		return -EINVAL;
 884
 885	rc = ntb_mw_get_align(nt->ndev, PIDX, num_mw, &xlat_align,
 886			      &xlat_align_size, NULL);
 887	if (rc)
 888		return rc;
 889
 890	xlat_size = round_up(size, xlat_align_size);
 891	buff_size = round_up(size, xlat_align);
 892
 893	/* No need to re-setup */
 894	if (mw->xlat_size == xlat_size)
 895		return 0;
 896
 897	if (mw->buff_size)
 898		ntb_free_mw(nt, num_mw);
 899
 900	/* Alloc memory for receiving data.  Must be aligned */
 901	mw->xlat_size = xlat_size;
 902	mw->buff_size = buff_size;
 903	mw->alloc_size = buff_size;
 904
 905	rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
 906	if (rc) {
 907		mw->alloc_size *= 2;
 908		rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
 909		if (rc) {
 910			dev_err(&pdev->dev,
 911				"Unable to alloc aligned MW buff\n");
 912			mw->xlat_size = 0;
 913			mw->buff_size = 0;
 914			mw->alloc_size = 0;
 915			return rc;
 916		}
 917	}
 918
 919	/* Notify HW the memory location of the receive buffer */
 920	rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
 921			      mw->xlat_size);
 922	if (rc) {
 923		dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 924		ntb_free_mw(nt, num_mw);
 925		return -EIO;
 926	}
 927
 928	return 0;
 929}
 930
 931static void ntb_qp_link_context_reset(struct ntb_transport_qp *qp)
 932{
 933	qp->link_is_up = false;
 934	qp->active = false;
 935
 936	qp->tx_index = 0;
 937	qp->rx_index = 0;
 938	qp->rx_bytes = 0;
 939	qp->rx_pkts = 0;
 940	qp->rx_ring_empty = 0;
 941	qp->rx_err_no_buf = 0;
 942	qp->rx_err_oflow = 0;
 943	qp->rx_err_ver = 0;
 944	qp->rx_memcpy = 0;
 945	qp->rx_async = 0;
 946	qp->tx_bytes = 0;
 947	qp->tx_pkts = 0;
 948	qp->tx_ring_full = 0;
 949	qp->tx_err_no_buf = 0;
 950	qp->tx_memcpy = 0;
 951	qp->tx_async = 0;
 952}
 953
 954static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 955{
 956	ntb_qp_link_context_reset(qp);
 957	if (qp->remote_rx_info)
 958		qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 959}
 960
 961static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 962{
 963	struct ntb_transport_ctx *nt = qp->transport;
 964	struct pci_dev *pdev = nt->ndev->pdev;
 965
 966	dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 967
 968	cancel_delayed_work_sync(&qp->link_work);
 969	ntb_qp_link_down_reset(qp);
 970
 971	if (qp->event_handler)
 972		qp->event_handler(qp->cb_data, qp->link_is_up);
 973}
 974
 975static void ntb_qp_link_cleanup_work(struct work_struct *work)
 976{
 977	struct ntb_transport_qp *qp = container_of(work,
 978						   struct ntb_transport_qp,
 979						   link_cleanup);
 980	struct ntb_transport_ctx *nt = qp->transport;
 981
 982	ntb_qp_link_cleanup(qp);
 983
 984	if (nt->link_is_up)
 985		schedule_delayed_work(&qp->link_work,
 986				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 987}
 988
 989static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 990{
 991	schedule_work(&qp->link_cleanup);
 992}
 993
 994static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 995{
 996	struct ntb_transport_qp *qp;
 997	u64 qp_bitmap_alloc;
 998	unsigned int i, count;
 999
1000	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1001
1002	/* Pass along the info to any clients */
1003	for (i = 0; i < nt->qp_count; i++)
1004		if (qp_bitmap_alloc & BIT_ULL(i)) {
1005			qp = &nt->qp_vec[i];
1006			ntb_qp_link_cleanup(qp);
1007			cancel_work_sync(&qp->link_cleanup);
1008			cancel_delayed_work_sync(&qp->link_work);
1009		}
1010
1011	if (!nt->link_is_up)
1012		cancel_delayed_work_sync(&nt->link_work);
1013
1014	for (i = 0; i < nt->mw_count; i++)
1015		ntb_free_mw(nt, i);
1016
1017	/* The scratchpad registers keep the values if the remote side
1018	 * goes down, blast them now to give them a sane value the next
1019	 * time they are accessed
1020	 */
1021	count = ntb_spad_count(nt->ndev);
1022	for (i = 0; i < count; i++)
1023		ntb_spad_write(nt->ndev, i, 0);
1024}
1025
1026static void ntb_transport_link_cleanup_work(struct work_struct *work)
1027{
1028	struct ntb_transport_ctx *nt =
1029		container_of(work, struct ntb_transport_ctx, link_cleanup);
1030
1031	guard(mutex)(&nt->link_event_lock);
1032	ntb_transport_link_cleanup(nt);
1033}
1034
1035static void ntb_transport_event_callback(void *data)
1036{
1037	struct ntb_transport_ctx *nt = data;
1038
1039	if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
1040		schedule_delayed_work(&nt->link_work, 0);
1041	else
1042		schedule_work(&nt->link_cleanup);
1043}
1044
1045static void ntb_transport_link_work(struct work_struct *work)
1046{
1047	struct ntb_transport_ctx *nt =
1048		container_of(work, struct ntb_transport_ctx, link_work.work);
1049	struct ntb_dev *ndev = nt->ndev;
1050	struct pci_dev *pdev = ndev->pdev;
1051	resource_size_t size;
1052	u32 val;
1053	int rc = 0, i, spad;
1054
1055	guard(mutex)(&nt->link_event_lock);
1056
1057	/* send the local info, in the opposite order of the way we read it */
1058
1059	if (nt->use_msi) {
1060		rc = ntb_msi_setup_mws(ndev);
1061		if (rc) {
1062			dev_warn(&pdev->dev,
1063				 "Failed to register MSI memory window: %d\n",
1064				 rc);
1065			nt->use_msi = false;
1066		}
1067	}
1068
1069	for (i = 0; i < nt->qp_count; i++)
1070		ntb_transport_setup_qp_msi(nt, i);
1071
1072	for (i = 0; i < nt->mw_count; i++) {
1073		size = nt->mw_vec[i].phys_size;
1074
1075		if (max_mw_size && size > max_mw_size)
1076			size = max_mw_size;
1077
1078		spad = MW0_SZ_HIGH + (i * 2);
1079		ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
1080
1081		spad = MW0_SZ_LOW + (i * 2);
1082		ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
1083	}
1084
1085	ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
1086
1087	ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
1088
1089	ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
1090
1091	/* Query the remote side for its info */
1092	val = ntb_spad_read(ndev, VERSION);
1093	dev_dbg(&pdev->dev, "Remote version = %d\n", val);
1094	if (val != NTB_TRANSPORT_VERSION)
1095		goto out;
1096
1097	val = ntb_spad_read(ndev, NUM_QPS);
1098	dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
1099	if (val != nt->qp_count)
1100		goto out;
1101
1102	val = ntb_spad_read(ndev, NUM_MWS);
1103	dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
1104	if (val != nt->mw_count)
1105		goto out;
1106
1107	for (i = 0; i < nt->mw_count; i++) {
1108		u64 val64;
1109
1110		val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
1111		val64 = (u64)val << 32;
1112
1113		val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
1114		val64 |= val;
1115
1116		dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
1117
1118		rc = ntb_set_mw(nt, i, val64);
1119		if (rc)
1120			goto out1;
1121	}
1122
1123	nt->link_is_up = true;
1124
1125	for (i = 0; i < nt->qp_count; i++) {
1126		struct ntb_transport_qp *qp = &nt->qp_vec[i];
1127
1128		ntb_transport_setup_qp_mw(nt, i);
1129		ntb_transport_setup_qp_peer_msi(nt, i);
1130
1131		if (qp->client_ready)
1132			schedule_delayed_work(&qp->link_work, 0);
1133	}
1134
1135	return;
1136
1137out1:
1138	for (i = 0; i < nt->mw_count; i++)
1139		ntb_free_mw(nt, i);
1140
1141	/* if there's an actual failure, we should just bail */
1142	if (rc < 0)
1143		return;
1144
1145out:
1146	if (ntb_link_is_up(ndev, NULL, NULL) == 1)
1147		schedule_delayed_work(&nt->link_work,
1148				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1149}
1150
1151static void ntb_qp_link_work(struct work_struct *work)
1152{
1153	struct ntb_transport_qp *qp = container_of(work,
1154						   struct ntb_transport_qp,
1155						   link_work.work);
1156	struct pci_dev *pdev = qp->ndev->pdev;
1157	struct ntb_transport_ctx *nt = qp->transport;
1158	int val;
1159
1160	WARN_ON(!nt->link_is_up);
1161
1162	val = ntb_spad_read(nt->ndev, QP_LINKS);
1163
1164	ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
1165
1166	/* query remote spad for qp ready bits */
1167	dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
1168
1169	/* See if the remote side is up */
1170	if (val & BIT(qp->qp_num)) {
1171		dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
1172		qp->link_is_up = true;
1173		qp->active = true;
1174
1175		if (qp->event_handler)
1176			qp->event_handler(qp->cb_data, qp->link_is_up);
1177
1178		if (qp->active)
1179			tasklet_schedule(&qp->rxc_db_work);
1180	} else if (nt->link_is_up)
1181		schedule_delayed_work(&qp->link_work,
1182				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1183}
1184
1185static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
1186				    unsigned int qp_num)
1187{
1188	struct ntb_transport_qp *qp;
1189	phys_addr_t mw_base;
1190	resource_size_t mw_size;
1191	unsigned int num_qps_mw, tx_size;
1192	unsigned int mw_num, mw_count, qp_count;
1193	u64 qp_offset;
1194
1195	mw_count = nt->mw_count;
1196	qp_count = nt->qp_count;
1197
1198	mw_num = QP_TO_MW(nt, qp_num);
1199
1200	qp = &nt->qp_vec[qp_num];
1201	qp->qp_num = qp_num;
1202	qp->transport = nt;
1203	qp->ndev = nt->ndev;
1204	qp->client_ready = false;
1205	qp->event_handler = NULL;
1206	ntb_qp_link_context_reset(qp);
1207
1208	if (mw_num < qp_count % mw_count)
1209		num_qps_mw = qp_count / mw_count + 1;
1210	else
1211		num_qps_mw = qp_count / mw_count;
1212
1213	mw_base = nt->mw_vec[mw_num].phys_addr;
1214	mw_size = nt->mw_vec[mw_num].phys_size;
1215
1216	if (max_mw_size && mw_size > max_mw_size)
1217		mw_size = max_mw_size;
1218
1219	tx_size = (unsigned int)mw_size / num_qps_mw;
1220	qp_offset = tx_size * (qp_num / mw_count);
1221
1222	qp->tx_mw_size = tx_size;
1223	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1224	if (!qp->tx_mw)
1225		return -EINVAL;
1226
1227	qp->tx_mw_phys = mw_base + qp_offset;
1228	if (!qp->tx_mw_phys)
1229		return -EINVAL;
1230
1231	tx_size -= sizeof(struct ntb_rx_info);
1232	qp->rx_info = qp->tx_mw + tx_size;
1233
1234	/* Due to housekeeping, there must be atleast 2 buffs */
1235	qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1236	qp->tx_max_entry = tx_size / qp->tx_max_frame;
1237
1238	if (nt->debugfs_node_dir) {
1239		char debugfs_name[8];
1240
1241		snprintf(debugfs_name, sizeof(debugfs_name), "qp%d", qp_num);
1242		qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1243						     nt->debugfs_node_dir);
1244
1245		qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1246							qp->debugfs_dir, qp,
1247							&ntb_qp_debugfs_stats);
1248	} else {
1249		qp->debugfs_dir = NULL;
1250		qp->debugfs_stats = NULL;
1251	}
1252
1253	INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1254	INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1255
1256	spin_lock_init(&qp->ntb_rx_q_lock);
1257	spin_lock_init(&qp->ntb_tx_free_q_lock);
1258
1259	INIT_LIST_HEAD(&qp->rx_post_q);
1260	INIT_LIST_HEAD(&qp->rx_pend_q);
1261	INIT_LIST_HEAD(&qp->rx_free_q);
1262	INIT_LIST_HEAD(&qp->tx_free_q);
1263
1264	tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1265		     (unsigned long)qp);
1266
1267	return 0;
1268}
1269
1270static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1271{
1272	struct ntb_transport_ctx *nt;
1273	struct ntb_transport_mw *mw;
1274	unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1275	u64 qp_bitmap;
1276	int node;
1277	int rc, i;
1278
1279	mw_count = ntb_peer_mw_count(ndev);
1280
1281	if (!ndev->ops->mw_set_trans) {
1282		dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1283		return -EINVAL;
1284	}
1285
1286	if (ntb_db_is_unsafe(ndev))
1287		dev_dbg(&ndev->dev,
1288			"doorbell is unsafe, proceed anyway...\n");
1289	if (ntb_spad_is_unsafe(ndev))
1290		dev_dbg(&ndev->dev,
1291			"scratchpad is unsafe, proceed anyway...\n");
1292
1293	if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1294		dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1295
1296	node = dev_to_node(&ndev->dev);
1297
1298	nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1299	if (!nt)
1300		return -ENOMEM;
1301
1302	nt->ndev = ndev;
1303
1304	/*
1305	 * If we are using MSI, and have at least one extra memory window,
1306	 * we will reserve the last MW for the MSI window.
1307	 */
1308	if (use_msi && mw_count > 1) {
1309		rc = ntb_msi_init(ndev, ntb_transport_msi_desc_changed);
1310		if (!rc) {
1311			mw_count -= 1;
1312			nt->use_msi = true;
1313		}
1314	}
1315
1316	spad_count = ntb_spad_count(ndev);
1317
1318	/* Limit the MW's based on the availability of scratchpads */
1319
1320	if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1321		nt->mw_count = 0;
1322		rc = -EINVAL;
1323		goto err;
1324	}
1325
1326	max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1327	nt->mw_count = min(mw_count, max_mw_count_for_spads);
1328
1329	nt->msi_spad_offset = nt->mw_count * 2 + MW0_SZ_HIGH;
1330
1331	nt->mw_vec = kcalloc_node(mw_count, sizeof(*nt->mw_vec),
1332				  GFP_KERNEL, node);
1333	if (!nt->mw_vec) {
1334		rc = -ENOMEM;
1335		goto err;
1336	}
1337
1338	for (i = 0; i < mw_count; i++) {
1339		mw = &nt->mw_vec[i];
1340
1341		rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1342					  &mw->phys_size);
1343		if (rc)
1344			goto err1;
1345
1346		mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1347		if (!mw->vbase) {
1348			rc = -ENOMEM;
1349			goto err1;
1350		}
1351
1352		mw->buff_size = 0;
1353		mw->xlat_size = 0;
1354		mw->virt_addr = NULL;
1355		mw->dma_addr = 0;
1356	}
1357
1358	qp_bitmap = ntb_db_valid_mask(ndev);
1359
1360	qp_count = ilog2(qp_bitmap);
1361	if (nt->use_msi) {
1362		qp_count -= 1;
1363		nt->msi_db_mask = BIT_ULL(qp_count);
1364		ntb_db_clear_mask(ndev, nt->msi_db_mask);
1365	}
1366
1367	if (max_num_clients && max_num_clients < qp_count)
1368		qp_count = max_num_clients;
1369	else if (nt->mw_count < qp_count)
1370		qp_count = nt->mw_count;
1371
1372	qp_bitmap &= BIT_ULL(qp_count) - 1;
1373
1374	nt->qp_count = qp_count;
1375	nt->qp_bitmap = qp_bitmap;
1376	nt->qp_bitmap_free = qp_bitmap;
1377
1378	nt->qp_vec = kcalloc_node(qp_count, sizeof(*nt->qp_vec),
1379				  GFP_KERNEL, node);
1380	if (!nt->qp_vec) {
1381		rc = -ENOMEM;
1382		goto err1;
1383	}
1384
1385	if (nt_debugfs_dir) {
1386		nt->debugfs_node_dir =
1387			debugfs_create_dir(pci_name(ndev->pdev),
1388					   nt_debugfs_dir);
1389	}
1390
1391	for (i = 0; i < qp_count; i++) {
1392		rc = ntb_transport_init_queue(nt, i);
1393		if (rc)
1394			goto err2;
1395	}
1396
1397	mutex_init(&nt->link_event_lock);
1398	INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1399	INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1400
1401	rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1402	if (rc)
1403		goto err2;
1404
1405	INIT_LIST_HEAD(&nt->client_devs);
1406	rc = ntb_bus_init(nt);
1407	if (rc)
1408		goto err3;
1409
1410	nt->link_is_up = false;
1411	ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1412	ntb_link_event(ndev);
1413
1414	return 0;
1415
1416err3:
1417	ntb_clear_ctx(ndev);
1418err2:
1419	kfree(nt->qp_vec);
1420err1:
1421	while (i--) {
1422		mw = &nt->mw_vec[i];
1423		iounmap(mw->vbase);
1424	}
1425	kfree(nt->mw_vec);
1426err:
1427	kfree(nt);
1428	return rc;
1429}
1430
1431static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1432{
1433	struct ntb_transport_ctx *nt = ndev->ctx;
1434	struct ntb_transport_qp *qp;
1435	u64 qp_bitmap_alloc;
1436	int i;
1437
1438	ntb_transport_link_cleanup(nt);
1439	cancel_work_sync(&nt->link_cleanup);
1440	cancel_delayed_work_sync(&nt->link_work);
1441
1442	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1443
1444	/* verify that all the qp's are freed */
1445	for (i = 0; i < nt->qp_count; i++) {
1446		qp = &nt->qp_vec[i];
1447		if (qp_bitmap_alloc & BIT_ULL(i))
1448			ntb_transport_free_queue(qp);
1449		debugfs_remove_recursive(qp->debugfs_dir);
1450	}
1451
1452	ntb_link_disable(ndev);
1453	ntb_clear_ctx(ndev);
1454
1455	ntb_bus_remove(nt);
1456
1457	for (i = nt->mw_count; i--; ) {
1458		ntb_free_mw(nt, i);
1459		iounmap(nt->mw_vec[i].vbase);
1460	}
1461
1462	kfree(nt->qp_vec);
1463	kfree(nt->mw_vec);
1464	kfree(nt);
1465}
1466
1467static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1468{
1469	struct ntb_queue_entry *entry;
1470	void *cb_data;
1471	unsigned int len;
1472	unsigned long irqflags;
1473
1474	spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1475
1476	while (!list_empty(&qp->rx_post_q)) {
1477		entry = list_first_entry(&qp->rx_post_q,
1478					 struct ntb_queue_entry, entry);
1479		if (!(entry->flags & DESC_DONE_FLAG))
1480			break;
1481
1482		entry->rx_hdr->flags = 0;
1483		iowrite32(entry->rx_index, &qp->rx_info->entry);
1484
1485		cb_data = entry->cb_data;
1486		len = entry->len;
1487
1488		list_move_tail(&entry->entry, &qp->rx_free_q);
1489
1490		spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1491
1492		if (qp->rx_handler && qp->client_ready)
1493			qp->rx_handler(qp, qp->cb_data, cb_data, len);
1494
1495		spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1496	}
1497
1498	spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1499}
1500
1501static void ntb_rx_copy_callback(void *data,
1502				 const struct dmaengine_result *res)
1503{
1504	struct ntb_queue_entry *entry = data;
1505
1506	/* we need to check DMA results if we are using DMA */
1507	if (res) {
1508		enum dmaengine_tx_result dma_err = res->result;
1509
1510		switch (dma_err) {
1511		case DMA_TRANS_READ_FAILED:
1512		case DMA_TRANS_WRITE_FAILED:
1513			entry->errors++;
1514			fallthrough;
1515		case DMA_TRANS_ABORTED:
1516		{
1517			struct ntb_transport_qp *qp = entry->qp;
1518			void *offset = qp->rx_buff + qp->rx_max_frame *
1519					qp->rx_index;
1520
1521			ntb_memcpy_rx(entry, offset);
1522			qp->rx_memcpy++;
1523			return;
1524		}
1525
1526		case DMA_TRANS_NOERROR:
1527		default:
1528			break;
1529		}
1530	}
1531
1532	entry->flags |= DESC_DONE_FLAG;
1533
1534	ntb_complete_rxc(entry->qp);
1535}
1536
1537static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1538{
1539	void *buf = entry->buf;
1540	size_t len = entry->len;
1541
1542	memcpy(buf, offset, len);
1543
1544	/* Ensure that the data is fully copied out before clearing the flag */
1545	wmb();
1546
1547	ntb_rx_copy_callback(entry, NULL);
1548}
1549
1550static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1551{
1552	struct dma_async_tx_descriptor *txd;
1553	struct ntb_transport_qp *qp = entry->qp;
1554	struct dma_chan *chan = qp->rx_dma_chan;
1555	struct dma_device *device;
1556	size_t pay_off, buff_off, len;
1557	struct dmaengine_unmap_data *unmap;
1558	dma_cookie_t cookie;
1559	void *buf = entry->buf;
1560
1561	len = entry->len;
1562	device = chan->device;
1563	pay_off = (size_t)offset & ~PAGE_MASK;
1564	buff_off = (size_t)buf & ~PAGE_MASK;
1565
1566	if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1567		goto err;
1568
1569	unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1570	if (!unmap)
1571		goto err;
1572
1573	unmap->len = len;
1574	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1575				      pay_off, len, DMA_TO_DEVICE);
1576	if (dma_mapping_error(device->dev, unmap->addr[0]))
1577		goto err_get_unmap;
1578
1579	unmap->to_cnt = 1;
1580
1581	unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1582				      buff_off, len, DMA_FROM_DEVICE);
1583	if (dma_mapping_error(device->dev, unmap->addr[1]))
1584		goto err_get_unmap;
1585
1586	unmap->from_cnt = 1;
1587
1588	txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1589					     unmap->addr[0], len,
1590					     DMA_PREP_INTERRUPT);
1591	if (!txd)
1592		goto err_get_unmap;
1593
1594	txd->callback_result = ntb_rx_copy_callback;
1595	txd->callback_param = entry;
1596	dma_set_unmap(txd, unmap);
1597
1598	cookie = dmaengine_submit(txd);
1599	if (dma_submit_error(cookie))
1600		goto err_set_unmap;
1601
1602	dmaengine_unmap_put(unmap);
1603
1604	qp->last_cookie = cookie;
1605
1606	qp->rx_async++;
1607
1608	return 0;
1609
1610err_set_unmap:
1611	dmaengine_unmap_put(unmap);
1612err_get_unmap:
1613	dmaengine_unmap_put(unmap);
1614err:
1615	return -ENXIO;
1616}
1617
1618static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1619{
1620	struct ntb_transport_qp *qp = entry->qp;
1621	struct dma_chan *chan = qp->rx_dma_chan;
1622	int res;
1623
1624	if (!chan)
1625		goto err;
1626
1627	if (entry->len < copy_bytes)
1628		goto err;
1629
1630	res = ntb_async_rx_submit(entry, offset);
1631	if (res < 0)
1632		goto err;
1633
1634	if (!entry->retries)
1635		qp->rx_async++;
1636
1637	return;
1638
1639err:
1640	ntb_memcpy_rx(entry, offset);
1641	qp->rx_memcpy++;
1642}
1643
1644static int ntb_process_rxc(struct ntb_transport_qp *qp)
1645{
1646	struct ntb_payload_header *hdr;
1647	struct ntb_queue_entry *entry;
1648	void *offset;
1649
1650	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1651	hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1652
1653	dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1654		qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1655
1656	if (!(hdr->flags & DESC_DONE_FLAG)) {
1657		dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1658		qp->rx_ring_empty++;
1659		return -EAGAIN;
1660	}
1661
1662	if (hdr->flags & LINK_DOWN_FLAG) {
1663		dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1664		ntb_qp_link_down(qp);
1665		hdr->flags = 0;
1666		return -EAGAIN;
1667	}
1668
1669	if (hdr->ver != (u32)qp->rx_pkts) {
1670		dev_dbg(&qp->ndev->pdev->dev,
1671			"version mismatch, expected %llu - got %u\n",
1672			qp->rx_pkts, hdr->ver);
1673		qp->rx_err_ver++;
1674		return -EIO;
1675	}
1676
1677	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1678	if (!entry) {
1679		dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1680		qp->rx_err_no_buf++;
1681		return -EAGAIN;
1682	}
1683
1684	entry->rx_hdr = hdr;
1685	entry->rx_index = qp->rx_index;
1686
1687	if (hdr->len > entry->len) {
1688		dev_dbg(&qp->ndev->pdev->dev,
1689			"receive buffer overflow! Wanted %d got %d\n",
1690			hdr->len, entry->len);
1691		qp->rx_err_oflow++;
1692
1693		entry->len = -EIO;
1694		entry->flags |= DESC_DONE_FLAG;
1695
1696		ntb_complete_rxc(qp);
1697	} else {
1698		dev_dbg(&qp->ndev->pdev->dev,
1699			"RX OK index %u ver %u size %d into buf size %d\n",
1700			qp->rx_index, hdr->ver, hdr->len, entry->len);
1701
1702		qp->rx_bytes += hdr->len;
1703		qp->rx_pkts++;
1704
1705		entry->len = hdr->len;
1706
1707		ntb_async_rx(entry, offset);
1708	}
1709
1710	qp->rx_index++;
1711	qp->rx_index %= qp->rx_max_entry;
1712
1713	return 0;
1714}
1715
1716static void ntb_transport_rxc_db(unsigned long data)
1717{
1718	struct ntb_transport_qp *qp = (void *)data;
1719	int rc, i;
1720
1721	dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1722		__func__, qp->qp_num);
1723
1724	/* Limit the number of packets processed in a single interrupt to
1725	 * provide fairness to others
1726	 */
1727	for (i = 0; i < qp->rx_max_entry; i++) {
1728		rc = ntb_process_rxc(qp);
1729		if (rc)
1730			break;
1731	}
1732
1733	if (i && qp->rx_dma_chan)
1734		dma_async_issue_pending(qp->rx_dma_chan);
1735
1736	if (i == qp->rx_max_entry) {
1737		/* there is more work to do */
1738		if (qp->active)
1739			tasklet_schedule(&qp->rxc_db_work);
1740	} else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1741		/* the doorbell bit is set: clear it */
1742		ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1743		/* ntb_db_read ensures ntb_db_clear write is committed */
1744		ntb_db_read(qp->ndev);
1745
1746		/* an interrupt may have arrived between finishing
1747		 * ntb_process_rxc and clearing the doorbell bit:
1748		 * there might be some more work to do.
1749		 */
1750		if (qp->active)
1751			tasklet_schedule(&qp->rxc_db_work);
1752	}
1753}
1754
1755static void ntb_tx_copy_callback(void *data,
1756				 const struct dmaengine_result *res)
1757{
1758	struct ntb_queue_entry *entry = data;
1759	struct ntb_transport_qp *qp = entry->qp;
1760	struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1761
1762	/* we need to check DMA results if we are using DMA */
1763	if (res) {
1764		enum dmaengine_tx_result dma_err = res->result;
1765
1766		switch (dma_err) {
1767		case DMA_TRANS_READ_FAILED:
1768		case DMA_TRANS_WRITE_FAILED:
1769			entry->errors++;
1770			fallthrough;
1771		case DMA_TRANS_ABORTED:
1772		{
1773			void __iomem *offset =
1774				qp->tx_mw + qp->tx_max_frame *
1775				entry->tx_index;
1776
1777			/* resubmit via CPU */
1778			ntb_memcpy_tx(entry, offset);
1779			qp->tx_memcpy++;
1780			return;
1781		}
1782
1783		case DMA_TRANS_NOERROR:
1784		default:
1785			break;
1786		}
1787	}
1788
1789	iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1790
1791	if (qp->use_msi)
1792		ntb_msi_peer_trigger(qp->ndev, PIDX, &qp->peer_msi_desc);
1793	else
1794		ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1795
1796	/* The entry length can only be zero if the packet is intended to be a
1797	 * "link down" or similar.  Since no payload is being sent in these
1798	 * cases, there is nothing to add to the completion queue.
1799	 */
1800	if (entry->len > 0) {
1801		qp->tx_bytes += entry->len;
1802
1803		if (qp->tx_handler)
1804			qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1805				       entry->len);
1806	}
1807
1808	ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1809}
1810
1811static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1812{
1813#ifdef copy_to_nontemporal
1814	/*
1815	 * Using non-temporal mov to improve performance on non-cached
1816	 * writes. This only works if __iomem is strictly memory-like,
1817	 * but that is the case on x86-64
1818	 */
1819	copy_to_nontemporal(offset, entry->buf, entry->len);
1820#else
1821	memcpy_toio(offset, entry->buf, entry->len);
1822#endif
1823
1824	/* Ensure that the data is fully copied out before setting the flags */
1825	wmb();
1826
1827	ntb_tx_copy_callback(entry, NULL);
1828}
1829
1830static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1831			       struct ntb_queue_entry *entry)
1832{
1833	struct dma_async_tx_descriptor *txd;
1834	struct dma_chan *chan = qp->tx_dma_chan;
1835	struct dma_device *device;
1836	size_t len = entry->len;
1837	void *buf = entry->buf;
1838	size_t dest_off, buff_off;
1839	struct dmaengine_unmap_data *unmap;
1840	dma_addr_t dest;
1841	dma_cookie_t cookie;
1842
1843	device = chan->device;
1844	dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
1845	buff_off = (size_t)buf & ~PAGE_MASK;
1846	dest_off = (size_t)dest & ~PAGE_MASK;
1847
1848	if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1849		goto err;
1850
1851	unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1852	if (!unmap)
1853		goto err;
1854
1855	unmap->len = len;
1856	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1857				      buff_off, len, DMA_TO_DEVICE);
1858	if (dma_mapping_error(device->dev, unmap->addr[0]))
1859		goto err_get_unmap;
1860
1861	unmap->to_cnt = 1;
1862
1863	txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1864					     DMA_PREP_INTERRUPT);
1865	if (!txd)
1866		goto err_get_unmap;
1867
1868	txd->callback_result = ntb_tx_copy_callback;
1869	txd->callback_param = entry;
1870	dma_set_unmap(txd, unmap);
1871
1872	cookie = dmaengine_submit(txd);
1873	if (dma_submit_error(cookie))
1874		goto err_set_unmap;
1875
1876	dmaengine_unmap_put(unmap);
1877
1878	dma_async_issue_pending(chan);
1879
1880	return 0;
1881err_set_unmap:
1882	dmaengine_unmap_put(unmap);
1883err_get_unmap:
1884	dmaengine_unmap_put(unmap);
1885err:
1886	return -ENXIO;
1887}
1888
1889static void ntb_async_tx(struct ntb_transport_qp *qp,
1890			 struct ntb_queue_entry *entry)
1891{
1892	struct ntb_payload_header __iomem *hdr;
1893	struct dma_chan *chan = qp->tx_dma_chan;
1894	void __iomem *offset;
1895	int res;
1896
1897	entry->tx_index = qp->tx_index;
1898	offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1899	hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1900	entry->tx_hdr = hdr;
1901
1902	iowrite32(entry->len, &hdr->len);
1903	iowrite32((u32)qp->tx_pkts, &hdr->ver);
1904
1905	if (!chan)
1906		goto err;
1907
1908	if (entry->len < copy_bytes)
1909		goto err;
1910
1911	res = ntb_async_tx_submit(qp, entry);
1912	if (res < 0)
1913		goto err;
1914
1915	if (!entry->retries)
1916		qp->tx_async++;
1917
1918	return;
1919
1920err:
1921	ntb_memcpy_tx(entry, offset);
1922	qp->tx_memcpy++;
1923}
1924
1925static int ntb_process_tx(struct ntb_transport_qp *qp,
1926			  struct ntb_queue_entry *entry)
1927{
1928	if (!ntb_transport_tx_free_entry(qp)) {
1929		qp->tx_ring_full++;
1930		return -EAGAIN;
1931	}
1932
1933	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1934		if (qp->tx_handler)
1935			qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1936
1937		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1938			     &qp->tx_free_q);
1939		return 0;
1940	}
1941
1942	ntb_async_tx(qp, entry);
1943
1944	qp->tx_index++;
1945	qp->tx_index %= qp->tx_max_entry;
1946
1947	qp->tx_pkts++;
1948
1949	return 0;
1950}
1951
1952static void ntb_send_link_down(struct ntb_transport_qp *qp)
1953{
1954	struct pci_dev *pdev = qp->ndev->pdev;
1955	struct ntb_queue_entry *entry;
1956	int i, rc;
1957
1958	if (!qp->link_is_up)
1959		return;
1960
1961	dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1962
1963	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1964		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1965		if (entry)
1966			break;
1967		msleep(100);
1968	}
1969
1970	if (!entry)
1971		return;
1972
1973	entry->cb_data = NULL;
1974	entry->buf = NULL;
1975	entry->len = 0;
1976	entry->flags = LINK_DOWN_FLAG;
1977
1978	rc = ntb_process_tx(qp, entry);
1979	if (rc)
1980		dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1981			qp->qp_num);
1982
1983	ntb_qp_link_down_reset(qp);
1984}
1985
1986static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1987{
1988	return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1989}
1990
1991/**
1992 * ntb_transport_create_queue - Create a new NTB transport layer queue
1993 * @data: pointer for callback data
1994 * @client_dev: &struct device pointer
1995 * @handlers: pointer to various ntb queue (callback) handlers
1996 *
1997 * Create a new NTB transport layer queue and provide the queue with a callback
1998 * routine for both transmit and receive.  The receive callback routine will be
1999 * used to pass up data when the transport has received it on the queue.   The
2000 * transmit callback routine will be called when the transport has completed the
2001 * transmission of the data on the queue and the data is ready to be freed.
2002 *
2003 * RETURNS: pointer to newly created ntb_queue, NULL on error.
2004 */
2005struct ntb_transport_qp *
2006ntb_transport_create_queue(void *data, struct device *client_dev,
2007			   const struct ntb_queue_handlers *handlers)
2008{
2009	struct ntb_dev *ndev;
2010	struct pci_dev *pdev;
2011	struct ntb_transport_ctx *nt;
2012	struct ntb_queue_entry *entry;
2013	struct ntb_transport_qp *qp;
2014	u64 qp_bit;
2015	unsigned int free_queue;
2016	dma_cap_mask_t dma_mask;
2017	int node;
2018	int i;
2019
2020	ndev = dev_ntb(client_dev->parent);
2021	pdev = ndev->pdev;
2022	nt = ndev->ctx;
2023
2024	node = dev_to_node(&ndev->dev);
2025
2026	free_queue = ffs(nt->qp_bitmap_free);
2027	if (!free_queue)
2028		goto err;
2029
2030	/* decrement free_queue to make it zero based */
2031	free_queue--;
2032
2033	qp = &nt->qp_vec[free_queue];
2034	qp_bit = BIT_ULL(qp->qp_num);
2035
2036	nt->qp_bitmap_free &= ~qp_bit;
2037
2038	qp->cb_data = data;
2039	qp->rx_handler = handlers->rx_handler;
2040	qp->tx_handler = handlers->tx_handler;
2041	qp->event_handler = handlers->event_handler;
2042
2043	dma_cap_zero(dma_mask);
2044	dma_cap_set(DMA_MEMCPY, dma_mask);
2045
2046	if (use_dma) {
2047		qp->tx_dma_chan =
2048			dma_request_channel(dma_mask, ntb_dma_filter_fn,
2049					    (void *)(unsigned long)node);
2050		if (!qp->tx_dma_chan)
2051			dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
2052
2053		qp->rx_dma_chan =
2054			dma_request_channel(dma_mask, ntb_dma_filter_fn,
2055					    (void *)(unsigned long)node);
2056		if (!qp->rx_dma_chan)
2057			dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
2058	} else {
2059		qp->tx_dma_chan = NULL;
2060		qp->rx_dma_chan = NULL;
2061	}
2062
2063	qp->tx_mw_dma_addr = 0;
2064	if (qp->tx_dma_chan) {
2065		qp->tx_mw_dma_addr =
2066			dma_map_resource(qp->tx_dma_chan->device->dev,
2067					 qp->tx_mw_phys, qp->tx_mw_size,
2068					 DMA_FROM_DEVICE, 0);
2069		if (dma_mapping_error(qp->tx_dma_chan->device->dev,
2070				      qp->tx_mw_dma_addr)) {
2071			qp->tx_mw_dma_addr = 0;
2072			goto err1;
2073		}
2074	}
2075
2076	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
2077		qp->tx_dma_chan ? "DMA" : "CPU");
2078
2079	dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
2080		qp->rx_dma_chan ? "DMA" : "CPU");
2081
2082	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
2083		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2084		if (!entry)
2085			goto err1;
2086
2087		entry->qp = qp;
2088		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
2089			     &qp->rx_free_q);
2090	}
2091	qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
2092
2093	for (i = 0; i < qp->tx_max_entry; i++) {
2094		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2095		if (!entry)
2096			goto err2;
2097
2098		entry->qp = qp;
2099		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2100			     &qp->tx_free_q);
2101	}
2102
2103	ntb_db_clear(qp->ndev, qp_bit);
2104	ntb_db_clear_mask(qp->ndev, qp_bit);
2105
2106	dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
2107
2108	return qp;
2109
2110err2:
2111	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2112		kfree(entry);
2113err1:
2114	qp->rx_alloc_entry = 0;
2115	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2116		kfree(entry);
2117	if (qp->tx_mw_dma_addr)
2118		dma_unmap_resource(qp->tx_dma_chan->device->dev,
2119				   qp->tx_mw_dma_addr, qp->tx_mw_size,
2120				   DMA_FROM_DEVICE, 0);
2121	if (qp->tx_dma_chan)
2122		dma_release_channel(qp->tx_dma_chan);
2123	if (qp->rx_dma_chan)
2124		dma_release_channel(qp->rx_dma_chan);
2125	nt->qp_bitmap_free |= qp_bit;
2126err:
2127	return NULL;
2128}
2129EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
2130
2131/**
2132 * ntb_transport_free_queue - Frees NTB transport queue
2133 * @qp: NTB queue to be freed
2134 *
2135 * Frees NTB transport queue
2136 */
2137void ntb_transport_free_queue(struct ntb_transport_qp *qp)
2138{
2139	struct pci_dev *pdev;
2140	struct ntb_queue_entry *entry;
2141	u64 qp_bit;
2142
2143	if (!qp)
2144		return;
2145
2146	pdev = qp->ndev->pdev;
2147
2148	qp->active = false;
2149
2150	if (qp->tx_dma_chan) {
2151		struct dma_chan *chan = qp->tx_dma_chan;
2152		/* Putting the dma_chan to NULL will force any new traffic to be
2153		 * processed by the CPU instead of the DAM engine
2154		 */
2155		qp->tx_dma_chan = NULL;
2156
2157		/* Try to be nice and wait for any queued DMA engine
2158		 * transactions to process before smashing it with a rock
2159		 */
2160		dma_sync_wait(chan, qp->last_cookie);
2161		dmaengine_terminate_all(chan);
2162
2163		dma_unmap_resource(chan->device->dev,
2164				   qp->tx_mw_dma_addr, qp->tx_mw_size,
2165				   DMA_FROM_DEVICE, 0);
2166
2167		dma_release_channel(chan);
2168	}
2169
2170	if (qp->rx_dma_chan) {
2171		struct dma_chan *chan = qp->rx_dma_chan;
2172		/* Putting the dma_chan to NULL will force any new traffic to be
2173		 * processed by the CPU instead of the DAM engine
2174		 */
2175		qp->rx_dma_chan = NULL;
2176
2177		/* Try to be nice and wait for any queued DMA engine
2178		 * transactions to process before smashing it with a rock
2179		 */
2180		dma_sync_wait(chan, qp->last_cookie);
2181		dmaengine_terminate_all(chan);
2182		dma_release_channel(chan);
2183	}
2184
2185	qp_bit = BIT_ULL(qp->qp_num);
2186
2187	ntb_db_set_mask(qp->ndev, qp_bit);
2188	tasklet_kill(&qp->rxc_db_work);
2189
2190	cancel_delayed_work_sync(&qp->link_work);
2191
2192	qp->cb_data = NULL;
2193	qp->rx_handler = NULL;
2194	qp->tx_handler = NULL;
2195	qp->event_handler = NULL;
2196
2197	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2198		kfree(entry);
2199
2200	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
2201		dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
2202		kfree(entry);
2203	}
2204
2205	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
2206		dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
2207		kfree(entry);
2208	}
2209
2210	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2211		kfree(entry);
2212
2213	qp->transport->qp_bitmap_free |= qp_bit;
2214
2215	dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
2216}
2217EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
2218
2219/**
2220 * ntb_transport_rx_remove - Dequeues enqueued rx packet
2221 * @qp: NTB queue to be freed
2222 * @len: pointer to variable to write enqueued buffers length
2223 *
2224 * Dequeues unused buffers from receive queue.  Should only be used during
2225 * shutdown of qp.
2226 *
2227 * RETURNS: NULL error value on error, or void* for success.
2228 */
2229void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
2230{
2231	struct ntb_queue_entry *entry;
2232	void *buf;
2233
2234	if (!qp || qp->client_ready)
2235		return NULL;
2236
2237	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
2238	if (!entry)
2239		return NULL;
2240
2241	buf = entry->cb_data;
2242	*len = entry->len;
2243
2244	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2245
2246	return buf;
2247}
2248EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2249
2250/**
2251 * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2252 * @qp: NTB transport layer queue the entry is to be enqueued on
2253 * @cb: per buffer pointer for callback function to use
2254 * @data: pointer to data buffer that incoming packets will be copied into
2255 * @len: length of the data buffer
2256 *
2257 * Enqueue a new receive buffer onto the transport queue into which a NTB
2258 * payload can be received into.
2259 *
2260 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2261 */
2262int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2263			     unsigned int len)
2264{
2265	struct ntb_queue_entry *entry;
2266
2267	if (!qp)
2268		return -EINVAL;
2269
2270	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2271	if (!entry)
2272		return -ENOMEM;
2273
2274	entry->cb_data = cb;
2275	entry->buf = data;
2276	entry->len = len;
2277	entry->flags = 0;
2278	entry->retries = 0;
2279	entry->errors = 0;
2280	entry->rx_index = 0;
2281
2282	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2283
2284	if (qp->active)
2285		tasklet_schedule(&qp->rxc_db_work);
2286
2287	return 0;
2288}
2289EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2290
2291/**
2292 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2293 * @qp: NTB transport layer queue the entry is to be enqueued on
2294 * @cb: per buffer pointer for callback function to use
2295 * @data: pointer to data buffer that will be sent
2296 * @len: length of the data buffer
2297 *
2298 * Enqueue a new transmit buffer onto the transport queue from which a NTB
2299 * payload will be transmitted.  This assumes that a lock is being held to
2300 * serialize access to the qp.
2301 *
2302 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2303 */
2304int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2305			     unsigned int len)
2306{
2307	struct ntb_queue_entry *entry;
2308	int rc;
2309
2310	if (!qp || !len)
2311		return -EINVAL;
2312
2313	/* If the qp link is down already, just ignore. */
2314	if (!qp->link_is_up)
2315		return 0;
2316
2317	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2318	if (!entry) {
2319		qp->tx_err_no_buf++;
2320		return -EBUSY;
2321	}
2322
2323	entry->cb_data = cb;
2324	entry->buf = data;
2325	entry->len = len;
2326	entry->flags = 0;
2327	entry->errors = 0;
2328	entry->retries = 0;
2329	entry->tx_index = 0;
2330
2331	rc = ntb_process_tx(qp, entry);
2332	if (rc)
2333		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2334			     &qp->tx_free_q);
2335
2336	return rc;
2337}
2338EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2339
2340/**
2341 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2342 * @qp: NTB transport layer queue to be enabled
2343 *
2344 * Notify NTB transport layer of client readiness to use queue
2345 */
2346void ntb_transport_link_up(struct ntb_transport_qp *qp)
2347{
2348	if (!qp)
2349		return;
2350
2351	qp->client_ready = true;
2352
2353	if (qp->transport->link_is_up)
2354		schedule_delayed_work(&qp->link_work, 0);
2355}
2356EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2357
2358/**
2359 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2360 * @qp: NTB transport layer queue to be disabled
2361 *
2362 * Notify NTB transport layer of client's desire to no longer receive data on
2363 * transport queue specified.  It is the client's responsibility to ensure all
2364 * entries on queue are purged or otherwise handled appropriately.
2365 */
2366void ntb_transport_link_down(struct ntb_transport_qp *qp)
2367{
2368	int val;
2369
2370	if (!qp)
2371		return;
2372
2373	qp->client_ready = false;
2374
2375	val = ntb_spad_read(qp->ndev, QP_LINKS);
2376
2377	ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2378
2379	if (qp->link_is_up)
2380		ntb_send_link_down(qp);
2381	else
2382		cancel_delayed_work_sync(&qp->link_work);
2383}
2384EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2385
2386/**
2387 * ntb_transport_link_query - Query transport link state
2388 * @qp: NTB transport layer queue to be queried
2389 *
2390 * Query connectivity to the remote system of the NTB transport queue
2391 *
2392 * RETURNS: true for link up or false for link down
2393 */
2394bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2395{
2396	if (!qp)
2397		return false;
2398
2399	return qp->link_is_up;
2400}
2401EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2402
2403/**
2404 * ntb_transport_qp_num - Query the qp number
2405 * @qp: NTB transport layer queue to be queried
2406 *
2407 * Query qp number of the NTB transport queue
2408 *
2409 * RETURNS: a zero based number specifying the qp number
2410 */
2411unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2412{
2413	if (!qp)
2414		return 0;
2415
2416	return qp->qp_num;
2417}
2418EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2419
2420/**
2421 * ntb_transport_max_size - Query the max payload size of a qp
2422 * @qp: NTB transport layer queue to be queried
2423 *
2424 * Query the maximum payload size permissible on the given qp
2425 *
2426 * RETURNS: the max payload size of a qp
2427 */
2428unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2429{
2430	unsigned int max_size;
2431	unsigned int copy_align;
2432	struct dma_chan *rx_chan, *tx_chan;
2433
2434	if (!qp)
2435		return 0;
2436
2437	rx_chan = qp->rx_dma_chan;
2438	tx_chan = qp->tx_dma_chan;
2439
2440	copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2441			 tx_chan ? tx_chan->device->copy_align : 0);
2442
2443	/* If DMA engine usage is possible, try to find the max size for that */
2444	max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2445	max_size = round_down(max_size, 1 << copy_align);
2446
2447	return max_size;
2448}
2449EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2450
2451unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2452{
2453	unsigned int head = qp->tx_index;
2454	unsigned int tail = qp->remote_rx_info->entry;
2455
2456	return tail >= head ? tail - head : qp->tx_max_entry + tail - head;
2457}
2458EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2459
2460static void ntb_transport_doorbell_callback(void *data, int vector)
2461{
2462	struct ntb_transport_ctx *nt = data;
2463	struct ntb_transport_qp *qp;
2464	u64 db_bits;
2465	unsigned int qp_num;
2466
2467	if (ntb_db_read(nt->ndev) & nt->msi_db_mask) {
2468		ntb_transport_msi_peer_desc_changed(nt);
2469		ntb_db_clear(nt->ndev, nt->msi_db_mask);
2470	}
2471
2472	db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2473		   ntb_db_vector_mask(nt->ndev, vector));
2474
2475	while (db_bits) {
2476		qp_num = __ffs(db_bits);
2477		qp = &nt->qp_vec[qp_num];
2478
2479		if (qp->active)
2480			tasklet_schedule(&qp->rxc_db_work);
2481
2482		db_bits &= ~BIT_ULL(qp_num);
2483	}
2484}
2485
2486static const struct ntb_ctx_ops ntb_transport_ops = {
2487	.link_event = ntb_transport_event_callback,
2488	.db_event = ntb_transport_doorbell_callback,
2489};
2490
2491static struct ntb_client ntb_transport_client = {
2492	.ops = {
2493		.probe = ntb_transport_probe,
2494		.remove = ntb_transport_free,
2495	},
2496};
2497
2498static int __init ntb_transport_init(void)
2499{
2500	int rc;
2501
2502	pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2503
2504	if (debugfs_initialized())
2505		nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2506
2507	rc = bus_register(&ntb_transport_bus);
2508	if (rc)
2509		goto err_bus;
2510
2511	rc = ntb_register_client(&ntb_transport_client);
2512	if (rc)
2513		goto err_client;
2514
2515	return 0;
2516
2517err_client:
2518	bus_unregister(&ntb_transport_bus);
2519err_bus:
2520	debugfs_remove_recursive(nt_debugfs_dir);
2521	return rc;
2522}
2523module_init(ntb_transport_init);
2524
2525static void __exit ntb_transport_exit(void)
2526{
2527	ntb_unregister_client(&ntb_transport_client);
2528	bus_unregister(&ntb_transport_bus);
2529	debugfs_remove_recursive(nt_debugfs_dir);
2530}
2531module_exit(ntb_transport_exit);